Abstract: In this paper，YOLO-V5，a target detection algorithm based on deep learning， was combined with DeepSORT，a multi- target tracking algorithm，to achieve real- time detection and statistics of pedestrian flow information at the platform level of subway stations. Firstly，to reduce the problem of false detection and missed detection caused by pedestrians' mutual occlusion，the traditional pedestrian whole-body detection was changed to pedestrian head and shoulder detection. Then，the ReID model in DeepSORT was trained to extract only the head and shoulder features of pedestrians，to reduce the problem of inaccurate counting caused by frequent switching of pedestrian IDs in the tracking process.Finally，the optimized pedestrian detection and tracking model was applied to the subway platform level passenger flow detection， and different passenger flow information was extracted and counted according to the actual application scenario. The results showed that the model could effectively detect the degree of platform congestion，and count the number of people going up and down at the entrance and exit of the platform. The accuracy rate was 86% and the average FPS was 35，which could meet the application requirements of realtime detection of passenger flow information.

Key words: YOLO-V5, object detection, DeepSORT, object tracking, passenger flow detection